Deep Learning-Based Classification of Student GPA Integrating Psychological and Family Factors in the Post-Pandemic Era

Abstract

Background: In the post-pandemic era, college students' academic performance is influenced by a range of non-cognitive factors, which often reduces the accuracy of conventional Grade Point Average (GPA) prediction models. For this, we developed a deep-learning–based GPA classification framework that integrates family background and psychological evaluation indicators, and empirically revealed the underlying associations among these dimensions. Methods: Data were collected from 1,692 undergraduates at a Chinese university. The dataset included family background factors such as gender, family economic situation, only-child, and left-behind years, as well as SCL-90 psychological evaluation scores and GPA records. Four deep learning models were evaluated: TabTransformer, DCNv2, AutoInt, and MLP-ResNet. In addition, a lightweight feature-gating mechanism was incorporated to improve feature selection in high-dimensional heterogeneous data. Model performance was evaluated using Accuracy and Area Under the ROC Curve (AUC). Associations among variables were analysed using Spearman's rank correlation, χ² tests, and t-SNE visualization. Results: The TabTransformer with the gating mechanism achieved the highest performance among the tested models, with an Accuracy of 0.798 and an AUC of 0.833. GPA was significantly negatively correlated with SCL-90 domains, including depression and anxiety. Additionally, unfavorable family background factors—such as lower family economic status and longer periods of being left behind—were correlated with poorer psychological assessment outcomes. Conclusion: This study developed a deep-learning framework using family background and psychological evaluation factors to classify GPA, support academic risk identification, and inform targeted academic assistance and psychological interventions.